Double-Layer Distributed and Integrated Fault Detection Strategy for Non-Gaussian Dynamic Industrial Systems

Area of Science:

• Industrial Systems Monitoring
• Data-Driven Fault Detection
• Statistical Process Control

Background:

• Large-scale industrial systems generate complex multisensor data with dynamic Gaussian and non-Gaussian characteristics.
• Traditional methods like Principal Component Analysis (PCA) are limited by assumptions of Gaussian distribution and data independence.
• Existing methods struggle with concurrent dynamic and mixed-distribution data, necessitating advanced fault detection techniques.

Purpose of the Study:

• To develop a robust fault detection method for large-scale industrial systems exhibiting concurrent dynamic and mixed Gaussian/non-Gaussian characteristics.
• To address the limitations of traditional methods in handling complex real-world industrial data.
• To improve the accuracy and reliability of fault detection in dynamic and heterogeneous data environments.

Main Methods:

• A double-layer distributed, data-driven strategy integrating Laplacian score weighting and Bayesian inference.
• First layer: Jarque-Bera test for classifying variables into Gaussian and non-Gaussian blocks.
• Second layer: Dynamic augmentation, K-means clustering for local similarity, and Laplace scoring for variable importance, followed by integrated Bayesian inference with detection performance weighting.

Main Results:

• The proposed double-layer strategy effectively combines diverse data characteristics: distribution differences, dynamism, local similarity, and variable importance.
• The integrated Bayesian inference emphasizes the contribution of local models, enhancing detection sensitivity.
• Validation on the Tennessee Eastman production system and a diesel engine system demonstrated the method's superiority over traditional approaches.

Conclusions:

• The proposed double-layer distributed strategy offers a significant advancement in fault detection for complex industrial systems.
• The method's ability to handle mixed distributions and dynamic behavior makes it suitable for real-world applications.
• This approach provides a more comprehensive and accurate tool for ensuring the safety and efficiency of industrial operations.